1 / 31

Observation error specification

Observation error specification. Gérald Desroziers Météo-France and CNRS. Outline. General framework Methods for estimating observation error statistics Diagnostic of observation error variances Diagnostic of observation error correlations

misae
Download Presentation

Observation error specification

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Observation error specification Gérald Desroziers Météo-France and CNRS

  2. Outline • General framework • Methods for estimating observation error statistics • Diagnostic of observation error variances • Diagnostic of observation error correlations • Observation error correlation specification in the assimilation • Conclusion

  3. General formalism • Statistical linear estimation : • xa = xb+ dx = xb+ K d = xb+ BHT (HBHT+R)-1 d, with d = yo – H(xb ), innovation, K,gain matrix, B et R, covariances of background and observation errors. • Solution of the variational problem • J(dx)= dxT B-1 dx + (d-H dx)T R-1(d-H dx). • Incremental formulation(Courtier et al, 1994): • J(x)=(x - xb)T B-1 (x - xb)+ (yo-H (x))T R-1(yo-H (x)).

  4. General formalism • Observation errors (Daley, 1993): • o = yo – H (xt) = yo – yt+ yt– H (xt),where yt is the true state equiv. of yo • = oi+ oH . • oi is the instrument error. • oH is a complex function of the • type of observation (in situ or integrated), • resolution of the state (representativeness error), • precision of the observation operator (satellite observation) ...

  5. Outline • General framework • Methods for estimating observation error statistics • Diagnostic of observation error variances • Diagnostic of observation error correlations • Observation error correlation specification in the assimilation • Conclusion

  6. Hollingsworth and Lönnberg method (From Bouttier and Courtier, ECMWF)

  7. A posteriori « Jmin » diagnostics • We should have • E[J(xa) ] = p, with • p = total number of observations. • (Bennett et al, 1993) • More precisely, for a sub-part of Jo : • E[Joi(xa) ] = pi – Tr(Ri-1/2 HiAHiT Ri-1/2 ), with • pi : number of observations associated with Joi, • Ri,Hi : associated error cov. matrix and obs. operator. • (Talagrand, 1999)

  8. A posteriori « Jmin » diagnostics: optimization of R • Normalization of Ri : soiRi • Coef. soi diagnosed with soi = E[Joi(xa)]/(E[Joi(xa)])opt • = E[Joi(xa)]/(pi–Tr(Ri-1/2HiAHiTRi-1/2)), (Desroziers and Ivanov, 2001; Chapnik et al, 2004; Desroziers et al 2009) • Equivalent to a Maximum-likelihood estimation (Dee, 1998) • f(d|s) = 1 / ((2p)p det(D(s))1/2 exp (-1/2 dTD(s)-1d), • where D(s) is the covariance matrix of parameters s. • Optimal parameters s are those that minimize the Log-likelihood • L(s) = -log ( f(d|s) ).

  9. yo doa d o xa a dab xb xt Diagnostics in observation space • d = yo – H (xb) • doa = yo – H (xa) • dab = H (xa) – H (xb) • E[doa dT] = R • E[dab dT] = HBHT • E[dab doaT] = HAHT • <, ’> = E[’T] (Desroziers et al, 2005)

  10. Diagnostics in observation space:practical implementation • For any subset i with pi observations, simply compute • (soi)2 = Sk=1,pi (yoik – yaik)(yoik – ybik) / pi . • Covariances between different observation errors can also be computed: • (Coi,j)2 = Sk=1,pi,j (yoik – yaik)(yojk – ybjk) / pi,j • inter-channel covariances, • spatial covariances ...

  11. Convergence: vodiag(vo) Idealized case: analysis on an equatorial circle (40 000km). votrue = 4. Lb = 300 km / L0= 0 km.

  12. Convergence: vodiag(vo) Idealized case: analysis on an equatorial circle (40 000km). votrue = 4. Lb = 300 km / L0= 200 km.

  13. Outline • General framework • Methods for estimating observation error statistics • Diagnostic of observation error variances • Diagnostic of observation error correlations • Observation error correlation specification in the assimilation • Conclusion

  14. Observation error standard-deviations Normalization of Ri : soiRi Coef. soi diagnosed with soi = E[Joi(xa)]/(E[Joi(xa)])opt. Normalization coefficients of oi in the French Arpège 4D-Var (Chapnik, et al, 2004; Buehner, 2005; Desroziers et al, 2009)

  15. Satellite error standard-deviations (Bormann et al, ECMWF, 2010)

  16. Outline • General framework • Methods for estimating observation error statistics • Diagnostic of observation error variances • Diagnostic of observation error correlations • Observation error correlation specification in the assimilation • Conclusion

  17. AMSU-A inter-channel error correlations (Bormann and Bauer, ECMWF, 2010; Bormann et al, ECMWF, 2011)

  18. IASI inter-channel error correlations (Bormann et al, ECMWF, 2011)

  19. AIRS inter-channel error correlations (Garand et al, Environment Canada, 2007)

  20. AMVs spatial error correlations (Bormann et al, ECMWF, 2003)

  21. SSM/I spatial error correlations Spatial error correlations for the F13 SSM/I (solid lines; black: clear sample; grey: cloudy sample) (Bormann et al, ECMWF, 2011)

  22. Doppler radar wind spatial error correlations Radial error correlation Rllo (r). (Xu et al, NOAA, 2007)

  23. Outline • General framework • Methods for estimating observation error statistics • Diagnostic of observation error variances • Diagnostic of observation error correlations • Observation error correlation specification in the assimilation • Conclusion

  24. Representation of inter-channel error correlations in R (Garand et al, Environment Canada, 2007)

  25. Representation ofspatial error correlations in R • Construction of a square-root correlation model for a block Ri = Si-1 CiSi-T of R with horizontal correlations (Si-1 normalization by standard-dev., Ci correlation matrix) • Ci = UiUiT • Ui = TiSi-1Gi1/2 , where • Giis the spectral (Hankel) transform of the correlation function, • Si-1 is the inverse spectral transform • (with a low, but sufficient resolution to represent the spatial correlation), • Tiis an interpolator to observation locations. • (Fisher and Radnoti, ECMWF, 2006)

  26. Representation ofspatial error correlations in R • Approximation of Ciby Ci= I + S1K (li,k- 1) vi,kvi,kT , where only a limited number K of eigenpairs (li,k, vi,k) of Ciis used. • The eigenpairs of Ci can be determined by a Lanczos algorithm. • Approximation of Ci-1 given by Ci-1 = I + S1K (1/li,k- 1) vi,kvi,kT . • Computation of effective normalized departures zeffi= Ci-1 zi = 1/aizi+ S1K (1/li,k- 1 /ai) vi,k (vi,kT zi), where ai is a parameter accounting for the truncation K, and zi are the departures normalized by Si . (Fisher and Radnoti, ECMWF, 2006; Isaksen and Radnoti, ECMWF, 2010).

  27. Benefit of representingspatial error correlations in R (similar to Liu and Rabier, Météo-France, 2002) Lb = 200 km Lo = 200 km b = o = 1

  28. Benefit of representingspatial error correlations in R Lb = 20 km Lo = 200 km b = o = 1

  29. Outline • General framework • Methods for estimating observation error statistics • Diagnostic of observation error variances • Diagnostic of observation error correlations • Observation error correlation specification in the assimilation • Conclusion

  30. Conclusion (I) • Observation errors are not explicitely known. • They can be inferred by a comparison with • other observations or with the background (innovations). • Available diagnostics of observation errors (variances and correlations), • but relying on explicit or implicit hypotheses. • Correlation of observation errors can be found in many datasets: • SYNOP time-correlations, • AIRS, IASI inter-channel correlations, • AMVs, SSM/I, radar spatial correlations.

  31. Conclusion (II) • Observation error correlations are often neglected, but with an empirical thinning and/or an inflation of error variance. • Correlations can be more or less easily taken into account. • A relevant formulation for spatial error correlation has been proposed • and implemented in a real size system (ECMWF). • Possible parametrization of R-1 , with the addition of gradient observations (Brankart et al, 2009). • Correlated observations are less informative than uncorrelated observations, if Lo becomes close to Lb, even if R is well specified. • Opposite situation, if Lo >> Lb !

More Related